MCP server for LinkedIn automation with Claude Desktop. Handles login, job searches, and profile viewing with anti-detection measures.
This LinkedIn automation server has serious definition quality gaps. While all 4 tools are explicitly registered with basic schemas, descriptions are generic, parameter descriptions are minimal or absent, and output schemas are undocumented. The server uses Puppeteer for web automation, which is inherently fragile and does not align with API-first MCP patterns. No input validation, error guidance, or output shaping is visible. The code shows authentication via credentials (email/password) passed as parameters, a critical security anti-pattern. Tool naming is reasonable (verb-based: login, searchJobs, applyToJob, viewProfile) but descriptions lack the LLM-optimized context needed for robust agent selection and error recovery.
Applies to a specific job on LinkedIn. Attempts to auto-fill the application form with provided information.
Logs into LinkedIn with provided email and password. Handles session restoration from saved cookies and security checkpoints.
Searches for jobs on LinkedIn with optional filters. Extracts job listings including title, company, location, salary, and job description.
Views a LinkedIn profile by username and extracts profile information including headline, about, experience, and education.
CRITICAL SECURITY: Credentials (email, password) exposed as tool parameters. Agents will log these in traces and prompt history. Violates pattern:secret-injection. Must use environment variables or vault injection instead.
No output schemas documented. LLMs cannot predict what fields login/searchJobs/applyToJob/viewProfile return, preventing downstream tool chaining and forcing exploratory calls. This violates pattern:tool and wastes agent reasoning cycles.
Parameter descriptions are incomplete or missing for critical inputs. E.g. searchJobs.filters has type 'object' with no guidance on valid keys/values beyond brief inline hints (e.g. 'r86400 for 24h'). applyToJob.resumeUrl lacks format guidance (URL? local path?). LLMs cannot reliably construct valid payloads.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 41 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 0 | - | v1 |
No error handling guidance visible in code. login() returns JSON with success/message/needsManualVerification, but no actionable recovery hints. When a job application fails (network, rate limit, missing field), agents have no guidance on whether to retry, ask the user, or escalate.
Web scraping via Puppeteer is brittle and violates LinkedIn's Terms of Service. No consent mechanism, no legal disclaimer in tool descriptions. Agents could trigger IP bans, account lockouts, or legal action without knowing the risk.
Tool descriptions are vague. 'Searches for jobs on LinkedIn with optional filters' and 'Applies to a specific job' do not explain WHEN to use these tools, what dependencies exist (e.g. login required), or what the LLM should do if a search returns 0 results.
No pagination support visible in searchJobs. LinkedIn job searches can return hundreds/thousands of results. Code does not show limit/offset/cursor handling or documentation of result cap, risking context window exhaustion and silent data loss.
No idempotency guarantees. applyToJob has no deduplication, dry-run, or confirmation step. If an agent retries after a network timeout, the job application may be submitted twice, wasting the user's applications or triggering LinkedIn's duplicate-detection.
viewProfile parameter 'username' accepts 'LinkedIn username or profile URL path' but no format constraint is documented. Should be: 'username (no special chars) or full profile URL (e.g. linkedin.com/in/john-doe)'. Ambiguity invites LLM hallucination.